COLUMN: DeepSeek V4 Pro — When China Open-Sources AI and Shakes Silicon Valley

COLUMN: DeepSeek V4 Pro — When China Open-Sources AI and Shakes Silicon Valley

Introduction: The Bombshell From Hangzhou That Nobody Saw Coming

April 24, 2026: the day the AI market was repriced

On April 24, 2026, DeepSeek — a Chinese artificial intelligence laboratory based in Hangzhou, founded by hedge fund manager Liang Wenfeng — quietly published two new models on Hugging Face under an MIT license: DeepSeek V4 Pro and DeepSeek V4-Flash. No press conference. No product launch event. Just a model card, a technical report, and a GitHub repository. By the end of the week, Wall Street had shaved billions off the valuations of Nvidia, OpenAI's private investors were recalculating assumptions, and the AI industry was asking itself, once again, how a Chinese lab had managed to do what American companies charge twenty to thirty times more to do.

DeepSeek V4 Pro is a Mixture-of-Experts (MoE) model with 1.6 trillion total parameters — a number that sounds astronomical until you realize that only 49 billion parameters are activated per token during inference. This architecture is the engineering key to the model's economics: you get frontier-level performance at a fraction of the computational cost. Input pricing at launch: $0.435 per million tokens. Compare that to Claude Opus 4.7 at $25 per million tokens, or GPT-5.5 at $30 per million tokens. The math is not subtle. DeepSeek V4 Pro costs roughly sixty times less to run than its closest American competitors.

The benchmarks that triggered the shockwave

The numbers on the benchmark leaderboards landed like a series of detonations. On SWE-bench Verified — the standard measure of a model's ability to solve real-world software engineering tasks — DeepSeek V4 Pro scored 80.6%. On LiveCodeBench, a competitive programming benchmark evaluated on live problems, it reached 93.5% — a new record at the time of release. On Codeforces, the competitive programming platform where human experts hold Elo ratings like chess grandmasters, the model achieved an Elo of 3206, placing it in a category of performance previously associated with the very best human competitive programmers.

These are not marginal improvements over existing models. They represent a qualitative threshold: a model that can engage with software engineering tasks at a level that meaningfully competes with — and in specific domains exceeds — the capabilities of American frontier models that cost two orders of magnitude more to use. The technical community absorbed this relatively quickly. The financial community took a few days longer. The geopolitical community is still working through the implications.

The Technical Architecture: Why MoE Changes Everything

1.6 trillion parameters, 49 billion active: the Mixture-of-Experts logic

To understand why DeepSeek V4 Pro can price its API at $0.435 per million tokens while delivering frontier performance, you need to understand the Mixture-of-Experts (MoE) architecture. A traditional dense neural network — like the original GPT models — activates all of its parameters for every input token. As models scaled to hundreds of billions of parameters, the computational cost of this approach became the primary constraint on both training and inference. MoE offers a different design: the model has many specialized subnetworks (« experts »), and for each input, a routing mechanism activates only a small subset of them.

In DeepSeek V4 Pro's case: 1.6 trillion total parameters, but only 49 billion active per token. The model has the representational capacity of a 1.6T dense model but the inference cost of a roughly 49B dense model. This is not a new idea — Google, Meta, and Mistral have all explored MoE architectures — but DeepSeek has pushed the scaling and training of MoE to a new frontier. The result is a model whose economic profile breaks the assumed relationship between capability and cost that had structured AI market pricing since the GPT-3 era.

The context window and attention mechanisms

Beyond the MoE architecture, DeepSeek V4 Pro introduced significant innovations in attention mechanisms. The model operates with a 1 million token context window — enabling it to process and reason over book-length documents, large codebases, or extended multi-turn conversations without losing coherence. This is achieved through a combination of Compressed Sparse Attention (CSA) and Hierarchical Context Attention (HCA) mechanisms that make long-context processing computationally tractable at scale.

A 1 million token context window changes the practical use case profile of the model significantly. Legal document review across an entire contract portfolio. Codebase analysis across hundreds of files. Scientific literature synthesis across dozens of papers simultaneously. These are tasks that previous context windows made impractical not because the model lacked the capability but because the context budget ran out. The combination of MoE efficiency and long context makes V4 Pro an unusually versatile tool across enterprise applications.

DeepSeek V4-Flash: The Efficient Companion

284 billion parameters, 13 billion active: the economics of the smaller sibling

Released simultaneously with V4 Pro, DeepSeek V4-Flash targets a different market position. With 284 billion total parameters and 13 billion active per token, it is designed for applications where latency and cost are the primary constraints — real-time conversational AI, high-volume API calls, embedded model deployment. Priced at $0.14 per million input tokens and $0.28 per million output tokens, it undercuts even the current generation of « efficient » models from American providers.

The existence of both models — a frontier capability model and a high-efficiency model released simultaneously — signals a mature product strategy from DeepSeek. Rather than releasing a single model that tries to serve all use cases, the lab has deployed a portfolio approach that mirrors what Anthropic does with its Haiku/Sonnet/Opus lineup and what Google does with its Flash/Pro/Ultra tiers. This market sophistication from a lab that most Western analysts were dismissing two years ago as a derivative operation is itself a data point worth taking seriously.

Pricing made permanent in May 2026

One of the more significant signals to emerge from the months following the April launch was DeepSeek's confirmation on May 22, 2026 that the launch pricing for both models would be made permanent — not a promotional introductory rate. This matters because API pricing in the AI market has historically been fluid: companies have launched at low prices and raised them as adoption locked in users and raised switching costs. DeepSeek's commitment to permanent launch pricing is a competitive signal: it is not using price as a temporary acquisition tool. It is structuring its business model around low-cost high-volume access.

The implications for the AI market are significant. If frontier-level capabilities are available at permanently low prices from a Chinese open-source provider, the business model of American AI companies that rely on premium API pricing as their primary revenue source comes under structural pressure. This is not a temporary disruption. It is a repricing of what frontier AI capability costs — and it is permanent.

Trained on Huawei Ascend 950: The Chip Independence Story

How DeepSeek built frontier AI without Nvidia

DeepSeek V4 Pro was trained entirely on Huawei Ascend 950 chips — the latest generation of AI accelerators from Huawei's semiconductor division. This is not a footnote. It is one of the most consequential technical facts about the model, because it demonstrates that China has developed a viable domestic AI chip infrastructure capable of training frontier models without access to Nvidia's H100 or H200 GPUs — which have been subject to US export controls since 2022.

The Ascend 950 is not equivalent to the H200 on raw compute metrics. But DeepSeek's engineers have compensated for hardware limitations through algorithmic and systems-level optimizations that achieve competitive end-to-end training efficiency. This is a pattern that the Chinese technology industry has demonstrated repeatedly: when access to American technology is restricted, Chinese engineers do not simply fail — they find alternative approaches that, in some cases, develop into independently competitive capabilities. The chip war that Washington launched in 2022 has not stopped Chinese AI development. It has redirected it.

Nvidia's position in a world where DeepSeek trains on Huawei

The strategic implications for Nvidia are not catastrophic in the short term — the vast majority of AI training globally still runs on Nvidia hardware, and the company's position in the American and global markets remains dominant. But the narrative of indispensability has been dented. If China's leading AI lab can train a frontier model on domestic chips, the argument that export controls on Nvidia hardware will prevent Chinese AI advancement is demonstrably false. The export controls may slow certain development paths. They have not stopped the destination.

The financial markets understood this instinctively when DeepSeek V3 launched in January 2026 and triggered a sharp Nvidia selloff. The reaction to V4 Pro was more measured — partly because the market had already processed the structural implication, and partly because Nvidia's American and global business remains enormous. But the V4 Pro release confirmed the trajectory, not just the moment. Chinese AI development has achieved genuine chip independence at the frontier level.

The MIT License: Open Source as a Geopolitical Weapon

What open-sourcing a 1.6T model actually means

Releasing DeepSeek V4 Pro under an MIT license is not a neutral technical decision. It is a strategic choice with profound implications for the global AI landscape. The MIT license is the most permissive open-source license available — it allows anyone to use, modify, distribute, and build commercial products on top of the model without restriction, with only the requirement to preserve the original copyright notice. In practice: any developer, company, or government in the world can download, run, fine-tune, and deploy DeepSeek V4 Pro at no cost, with no usage restrictions, and without any contractual relationship with DeepSeek.

This creates a fundamentally different competitive dynamic from the closed-model approach of OpenAI, Anthropic, and Google DeepMind. A company that wants to build a product on top of GPT-5.5 must go through OpenAI's API, accept OpenAI's terms of service, and pay OpenAI's prices. A company that wants to build on DeepSeek V4 Pro can simply download the weights and run it on their own infrastructure. The economic and strategic leverage that American AI companies derive from their closed-model business model evaporates when a comparable open-source model exists.

The geopolitical dimension of open-source AI

China's choice to open-source its most capable frontier models — first DeepSeek V3 in January 2026, now V4 Pro — is a geopolitical strategy as much as a technical one. By releasing its best models as open source, China simultaneously: undermines the commercial moats of American AI companies, builds goodwill in the global developer community, accelerates the adoption of its own technical standards and architectures, and makes export control regimes largely irrelevant — you cannot embargo a model that anyone can download from Hugging Face.

It is worth noting the symmetry with China's hardware strategy. When the United States controls the hardware layer through export controls, China responds by making the software layer open. The result is a competitive landscape where China is selectively cooperative at the layer where it is strongest and where openness maximizes disruption for its competitors. This is not naive generosity. It is a carefully calibrated market strategy.

The National Intelligence Law: What Beijing Can Actually Demand

China's 2017 legislation and its implications for DeepSeek

Any serious assessment of DeepSeek must address China's National Intelligence Law of 2017, which requires Chinese companies and individuals to support, assist, and cooperate with national intelligence activities when requested. This law applies to DeepSeek as a Chinese company regardless of where it operates, where its models are deployed, or what its open-source licensing says. The question it raises: can data processed by DeepSeek's API be accessed by Chinese intelligence services? Can the model itself be required to include backdoors or information-reporting capabilities?

The honest answer is that the risk is real but hard to quantify precisely. The MIT-licensed weights that are publicly downloadable can be audited — and security researchers have begun doing so. The API service, run on DeepSeek's servers, operates under Chinese jurisdiction and is subject to the 2017 law. For enterprises handling sensitive data, the distinction between self-hosting the open-source weights versus using the DeepSeek API is therefore significant — the former avoids the data access risk, the latter does not.

The jailbreak controversy and the January 2026 incident

In January 2026, shortly after the DeepSeek V3 release, security researchers demonstrated that the model could be relatively easily jailbroken to produce content that its stated safety filters were designed to prevent — including, in some demonstrations, content related to sensitive Chinese political topics like Tiananmen Square that the model systematically refused to discuss in its default configuration. This controversy highlighted two distinct issues: the model's political content controls reflect Chinese government priorities, and the safety filters are technically less robust than those of comparable American models against targeted circumvention attempts.

The jailbreak controversy did not significantly slow DeepSeek's adoption — the capability profile was too compelling and the price differential too large for most commercial use cases where political content and safety vulnerabilities are not primary concerns. But it reinforced the perception that DeepSeek models require more careful deployment evaluation for applications in regulated industries or sensitive contexts. The MIT license means you can use these models freely — it does not mean the risk assessment is the same as for models built under American regulatory frameworks.

The Silicon Valley Response: Denial, Recalculation, and Adaptation

How American AI companies are responding

The responses from American AI companies to the DeepSeek V4 Pro release followed a now-familiar pattern. Initial communications from OpenAI and Anthropic emphasized proprietary capabilities and safety advantages. Analyst calls focused on the total addressable market argument: frontier AI is expanding the market faster than any single competitor can capture it, so the competitive pressure from DeepSeek may be less catastrophic than it appears. Venture capital firms with large positions in American AI companies put out notes arguing that American models maintain meaningful advantages in reasoning, alignment, and enterprise support ecosystems.

Some of these arguments have merit. Claude Opus 4.7's performance on complex reasoning tasks, particularly in areas requiring careful multi-step inference and nuanced understanding of context, remains ahead of DeepSeek V4 Pro on certain benchmarks. The enterprise support, compliance infrastructure, and legal indemnification that OpenAI and Anthropic offer have real value for large corporate customers. The American models are not simply worse than their Chinese competitor. They are differently positioned — and for many enterprise use cases, that positioning still commands a premium.

The API compatibility play: OpenAI's format adopted by DeepSeek

One of the more pointed competitive moves in the DeepSeek V4 Pro deployment is the model's explicit API compatibility with the OpenAI and Anthropic API formats. This means that any application already built to call the GPT-5.5 or Claude APIs can switch to DeepSeek V4 Pro by changing a single API endpoint URL and API key — no code rewriting required. The switching cost, which American companies have relied upon as a moat once developers are integrated, has been effectively zeroed out.

This is an elegant competitive move: rather than building a competing ecosystem, DeepSeek built an API that is a drop-in replacement for the existing ecosystem. Developers who want to cut their AI costs by a factor of sixty can do so without changing their code. The friction that protects American AI companies' installed base — the cost and effort of migration — has been eliminated by deliberate design. This is not accidental. It is a precision strike at the most defensible part of American AI companies' competitive position.

The Coding Benchmark Revolution: What SWE-bench 80.6% Really Means

Software engineering as the proving ground

The AI capability domain where DeepSeek V4 Pro's performance has had the most immediate practical impact is software engineering. The SWE-bench Verified score of 80.6% — measuring the model's ability to resolve real GitHub issues from popular open-source repositories — represents a threshold at which AI coding assistance moves from a useful productivity tool to something approaching an autonomous capability. At 80.6% on SWE-bench, the model can independently resolve most commonly encountered software engineering tasks without human intervention.

The practical implication is a significant compression in the cost of software development for teams that can effectively deploy and orchestrate AI coding agents. Individual developers using DeepSeek V4 Pro for coding assistance are reporting productivity multiples that were previously associated only with much more expensive model tiers. At $0.435 per million tokens, the economics of AI-augmented software development are no longer constrained by API costs for any organization of reasonable scale. The coding productivity revolution that was predicted when GPT-4 launched in 2023 is arriving — at a price point that makes it universally accessible.

Codeforces Elo 3206: the competitive programming frontier

The Codeforces Elo of 3206 achieved by DeepSeek V4 Pro deserves particular attention because the Codeforces platform measures performance against real human competitors in live competitions — not curated benchmark problems. An Elo of 3206 places the model in territory previously occupied by only a handful of the world's elite competitive programmers. This is not a synthetic benchmark that can be gamed by training on benchmark-adjacent data. It is a live performance measurement against the best human practitioners in algorithmics and competitive programming.

The implications extend beyond competitive programming as a domain. The reasoning patterns required for top-tier competitive programming — pattern recognition, algorithmic design, constraint optimization, edge case handling — are the same patterns that underlie many of the most valuable enterprise software engineering tasks. A model that performs at this level on competitive programming is almost certainly carrying over similar reasoning capabilities to real-world software tasks in ways that the SWE-bench score, comprehensive as it is, only partially captures.

Liang Wenfeng: The Founder Who Didn't Want to Be a Disruptor

The hedge fund manager who built a frontier AI lab

DeepSeek was founded by Liang Wenfeng, the co-founder of High-Flyer Capital Management, one of China's largest quantitative hedge funds. This origin is significant: DeepSeek was not a startup seeking venture capital validation or a national champion receiving government direction. It was founded by a mathematician and entrepreneur who built a quantitative trading firm, developed internal AI infrastructure for that firm, and then decided the technology warranted independent development as a standalone lab.

This origin story gives DeepSeek a different institutional culture from either American AI startups or Chinese state-directed technology projects. The lab's focus on research excellence, its willingness to release models openly without a commercial API-first strategy, and its engineering culture driven by academic rigor rather than product velocity reflect the priorities of a research institution funded by private capital rather than either venture returns or government direction. This does not resolve the National Intelligence Law question — no Chinese company is exempt from that framework — but it does suggest that DeepSeek's immediate strategic motivations are less directly statecraft-driven than some analysts have assumed.

The research culture that produced V4 Pro

The technical reports accompanying DeepSeek V4 Pro's release are remarkable documents. They are written with the transparency and rigor of academic papers, not marketing collateral. The architectural innovations — the MoE scaling decisions, the attention mechanism designs, the training efficiency optimizations — are explained with sufficient detail that the global research community can build on them. This openness is not naive: DeepSeek knows that sharing the architectural insights while retaining the trained weights and the production infrastructure still provides competitive advantage. But it reflects a genuine research culture that values contribution to the field alongside competitive positioning.

This culture has produced a laboratory that, in a remarkably short time, has gone from an internal infrastructure project at a hedge fund to the organization that published what is, by several metrics, the world's most capable open-source language model. The trajectory from DeepSeek V1 in 2023 to V4 Pro in April 2026 — three years, multiple frontier-level releases — is a pace of progress that the AI industry has not seen from any organization outside the American hyperscalers. It deserves to be taken seriously on its own terms.

The Broader AI Race: What DeepSeek V4 Pro Changes

The multipolar AI future is here

The release of DeepSeek V4 Pro is the clearest evidence yet that the AI landscape has become genuinely multipolar. For the first three years of the generative AI wave — roughly 2022 through 2025 — the frontier was effectively defined by three American organizations: OpenAI, Anthropic, and Google DeepMind. Other players contributed meaningfully to specific domains, but the overall capability frontier was American. DeepSeek V3 in January 2026 challenged that picture. V4 Pro confirms it: China now has a domestically developed, domestically trained, openly released frontier model that competes at the highest level of global AI capability.

The strategic implications of this multipolar reality are still being processed in Washington, in Brussels, and in the boardrooms of American AI companies. The export control strategy — restricting access to advanced semiconductors to maintain American AI superiority — has demonstrably failed to prevent China from reaching the frontier. The question is no longer whether China will develop frontier AI. The question is what policy frameworks, trade arrangements, and security protocols make sense in a world where the most capable open-source AI model is Chinese.

What American leadership in AI still means

The multipolar AI landscape does not mean American AI leadership is finished. OpenAI, Anthropic, and Google retain meaningful advantages in certain reasoning domains, in safety and alignment research, in enterprise integration, and in the regulatory and compliance infrastructure that large institutional customers require. The closed-model approach remains viable for high-trust applications where the MIT-license openness of DeepSeek creates risks that cannot be mitigated by self-hosting.

But the era in which American AI capability was so far ahead that Chinese alternatives were not serious competitive considerations is over. The competitive intensity that DeepSeek V4 Pro has introduced will likely accelerate the pace of American AI development — companies that face genuine competition at the frontier tend to move faster. In that sense, DeepSeek may be doing American AI more good than harm in the long run, even as it disrupts the current pricing structures. Competition is the engine of progress. Liang Wenfeng has ensured that American AI companies have a credible competitor. The field will be better for it.

The Regulation Question: What Brussels and Washington Should Do

European regulatory frameworks and open-source AI from China

The EU AI Act, which entered into force in August 2024, establishes a tiered regulatory framework based on risk levels. Open-source models — including DeepSeek V4 Pro — benefit from certain exemptions for non-commercial release, but these exemptions have limits: fine-tuned versions deployed for commercial purposes, or models modified for high-risk applications, fall under full regulatory scope. The European regulatory framework was designed primarily with American and European providers in mind. Its application to a Chinese open-source model with MIT licensing and a 1 million token context window operating across European cloud infrastructure is a genuinely novel regulatory question.

The key compliance questions for European enterprises deploying DeepSeek V4 Pro center on GDPR: data processed through DeepSeek's API is subject to Chinese jurisdiction under the National Intelligence Law, creating a direct conflict with GDPR's requirements on data transfers outside the EU. Self-hosted deployments of the open-source weights, running on EU infrastructure, sidestep the data transfer question — but create obligations around the model's AI Act classification that are still being worked through by national regulators.

The American response: new export controls or market adaptation?

In Washington, the response to DeepSeek V4 Pro has been a debate between two camps. One camp argues for tighter export controls — extending restrictions to cloud computing access, restricting Chinese AI companies from operating in American markets, and imposing security reviews on open-source models released by Chinese entities. The other camp argues that the export control approach has already demonstrated its limits and that the better response is to accelerate American AI investment, maintain the talent pipeline, and compete on capability rather than restriction.

The open-source release makes the restriction approach particularly difficult. You cannot export-control a model that is already on Hugging Face under an MIT license. The weights are public. They are being downloaded hundreds of thousands of times. Any serious restriction strategy would need to operate at the infrastructure layer — restricting deployment of Chinese-origin models on American cloud platforms — which raises significant legal and free speech questions. The market adaptation approach is harder politically but may be more practically effective. American AI companies competing with a sixty-times-cheaper competitor will either improve their efficiency or cede the market. Both outcomes are arguably better for the world than a protectionist framework that entrenches incumbency at the cost of innovation.

The Regulatory Response: Between Incomprehension and Urgency

Europe facing its own contradiction

The European Union finds itself in a paradoxical position. The European AI Act, which entered into force in 2024, imposes transparency and compliance obligations on "general-purpose" AI models deployed in the EU. DeepSeek V4 Pro, published under the MIT license and usable without notifying any authority, clearly falls into that category. But the AI Act's compliance mechanisms were designed primarily with large closed platforms in mind — OpenAI, Google, Anthropic — not with a Chinese open-weight model downloaded directly from Hugging Face by millions of European developers.

Europe also faces its own industrial contradiction. It has no competitive frontier proprietary model — neither Mistral nor any other European player operates at V4 Pro's scale. It therefore depends on American or Chinese actors for the most advanced AI capabilities. Adopting V4 Pro reduces dependence on the United States but creates a new dependence on China. European digital sovereignty is not served by this binary choice — it requires an industrial strategy that neither Brussels nor national capitals has yet formulated with sufficient urgency.

Washington searches for answers

On the American side, the political reaction was more immediate but not necessarily more coherent. Some members of Congress called for banning the use of DeepSeek in federal agencies — a measure already adopted for TikTok. Others proposed accelerating investment in domestic open-source AI, particularly through the National AI Research Resource (NAIRR) program. The Department of Commerce is examining whether Huawei's Ascend 950 chips should be added to the restricted list under export control regulations.

None of these responses is sufficient in isolation. The ban in federal agencies is a legitimate precautionary measure but does not address the fundamental question: how does the West maintain an advantage in AI when a Chinese open-source model rivals its best proprietary models at a fraction of the cost? The honest answer is that this advantage is narrowing. And the speed at which it narrows is accelerating.

DeepSeek and the AI Dimension of Global Competition

AI in the strategic geopolitical competition

The connection between DeepSeek and the broader geopolitical competition is not abstract. China, which maintains strategic ambiguity in the Russia-Ukraine conflict — neither Moscow's formal ally nor opposed to its aggression — is developing AI capabilities that could, in the near future, have dual applications. Advanced reasoning models like V4 Pro, trained on Huawei chips, exist within a technological ecosystem where the boundary between commercial and strategic use is deliberately blurred.

For Ukraine, which depends massively on Western technology to maintain its battlefield advantage, the implications are not immediate but merit attention. A world in which China has complete technological sovereignty in AI — training on domestic hardware, frontier models without American dependency — is a world in which Western leverage on Beijing is reduced. And the pressure on Beijing to cease economically supporting Russia's war machine is one of the central strategic tools available to the West.

The Western technological advantage: a strategic urgency

The West has not lost the AI race. GPT-5.4, Gemini 3.1 Pro, and Claude Opus 4.7 maintain measurable advantages on benchmarks of pure reasoning, multimodality, and general knowledge. But these advantages are narrowing faster than policymakers appear to realize. And in the long geopolitical competition — which is also an algorithmic competition — the question is not who wins which benchmark in April 2026, but who will have the most robust ecosystem, the deepest infrastructure, and the most resilient open-source community five years from now.

Maintaining that advantage demands massive investment — in fundamental research, in computing infrastructure, in talent development, in sovereign open-source capable of competing with DeepSeek on its own distribution strategy. It also demands serious policy thinking about what AI sovereignty means in the era of open-weight models: not simply who builds the best models, but who controls the infrastructure layers on which the global AI economy runs.

The Governance of Open Models: A Dangerous Regulatory Void

When transparency becomes a double-edged weapon

One of the most urgent debates that DeepSeek V4 Pro reignites is the governance of open AI models. The MIT license means that anyone — states, companies, malicious actors — can download the model, modify it, and deploy it for any use. This openness is a resource for the academic community and innovative startups. But it also represents a real risk: a model this powerful, available without restriction, can be fine-tuned to produce disinformation at scale, to automate social engineering attacks, or to generate malware with a sophistication previously inaccessible outside of well-funded state programs.

The European Union, with its AI Act progressively entering into force since 2024, attempts to frame general-purpose models, including open-source ones. But enforcing these rules on models hosted outside the EU remains a considerable challenge. DeepSeek V4 Pro, developed in China and distributed on Hugging Face from international servers, perfectly illustrates this regulatory gap: the model is available in Europe, used by European developers, subject to European law in theory, and impossible to effectively supervise in practice.

The precedent of DeepSeek V3 security vulnerabilities

It is important to recall that the previous version, DeepSeek V3, was the subject of a major security controversy in January 2026, when researchers demonstrated its vulnerability to "jailbreak" attacks allowing its ethical safeguards to be bypassed to generate potentially dangerous content. DeepSeek AI corrected some of these vulnerabilities, but concerns persisted. With V4 Pro, the question arises again: were the security tests sufficiently rigorous before public release? The MIT license places no obligation on users to report vulnerabilities or contribute security patches. The community is the safeguard. The community is not always sufficient.

These security questions are not theoretical. In a period of high geopolitical tensions — war in Ukraine, Sino-American technological rivalry — sophisticated AI models available without restriction constitute potential tools of informational warfare. The West must develop security frameworks for open models that allow benefiting from them while limiting the risks of malicious use. This is a difficult but not impossible equation. What is missing is not the technical capacity to build these frameworks. What is missing is the political will to prioritize them.

Conclusion: The Hangzhou Earthquake That Didn't Stop Shaking

What V4 Pro represents in the arc of AI development

When historians of technology look back at the AI decade of the 2020s, the DeepSeek V4 Pro release of April 2026 will be one of the markers they identify. Not because it was the most capable model ever released — that claim will be contested and surpassed many times over. But because it demonstrated, definitively and publicly, that the frontier of artificial intelligence is no longer a exclusively American domain, that frontier capability can be made available at prices that democratize access globally, and that the relationship between chip access, training efficiency, and model capability is not fixed by the hardware constraints that American export control policy assumed.

The AI race is more competitive, more global, and more open than the market structure of 2023 suggested it would be. That is a disruption for companies that had priced their services on the assumption of sustained scarcity. It is a windfall for every developer, researcher, and enterprise in the world that can now access frontier AI capabilities at a price that was unimaginable eighteen months ago. The earthquake from Hangzhou has not stopped shaking. The aftershocks will continue for years.

By Maxime Marquette, columnist

Columnist's transparency note

This column is based on publicly verifiable sources: DeepSeek's official technical reports and model cards published on Hugging Face, benchmark results published on SWE-bench, LiveCodeBench, and Codeforces, API pricing published on DeepSeek's official website, financial market data from Bloomberg and Reuters, and analyses published by The Verge, Ars Technica, MIT Technology Review, and various AI safety researchers. I hold no financial position in any AI company mentioned in this article. My assessment of the geopolitical dimensions of DeepSeek's open-source strategy reflects my own editorial judgment and should be read as informed analysis, not as a security or regulatory assessment. Enterprises evaluating DeepSeek deployment should obtain independent legal and compliance advice appropriate to their jurisdiction and use case.

Sources

Primary sources

Hugging Face — DeepSeek-V4-Pro model card and technical report — April 24, 2026

DeepSeek — V4 Pro official release announcement — April 24, 2026

SWE-bench — Verified leaderboard, DeepSeek V4 Pro results — April 2026

Codeforces — DeepSeek V4 Pro competitive programming rating — April 2026

Secondary sources

The Verge — DeepSeek V4 Pro shakes Silicon Valley again — April 24, 2026

Ars Technica — DeepSeek's latest model puts frontier AI under open license at prices American companies can't match — April 2026

MIT Technology Review — The Chinese AI model that changed what frontier means — April 2026

Foreign Policy — DeepSeek V4 Pro and the geopolitics of open-source AI — April 2026